> For the complete documentation index, see [llms.txt](https://docs.roboset.org/llms.txt). Markdown versions of documentation pages are available by appending `.md` to page URLs; this page is available as [Markdown](https://docs.roboset.org/training-platform/data-training.md).

# Data Training

### Training Modes

#### Reinforcement Learning (PPO)

Automated training uses PPO (Proximal Policy Optimization) from Stable Baselines3. PPO is the standard choice for continuous-control problems because it keeps policy updates stable across the high-dimensional action spaces robot arms require.

You choose an episode count and start the run. The policy learns through repeated trial and reward: it acts, the environment scores the action, and over many episodes the network's weights shift so higher-reward behavior becomes more likely. The UI streams current episode, total episodes, current reward, best reward, success rate, and training phase.

#### Manual Demonstration

You take direct control of the robot. Each command — movement, grab, release — steps the live environment, and the backend records the previous observation, the action, and the reward as a replayable trajectory. When you finish an attempt, the trajectory is saved.

Human demonstrations teach strategies far faster than random exploration. A person who already knows how to pick something up can show the system in seconds what RL might take thousands of episodes to discover.

#### Hybrid Learning (Behavior Cloning → RL)

When demonstrations exist for your task, training can begin with behavior cloning: the policy first learns to imitate your recorded actions, mapping observations to the actions you took. Once warm-started, PPO fine-tuning continues in the environment.

Behavior cloning followed by reinforcement learning is the same recipe frontier robotics labs use for real-world policies. Hybrid mode reduces wasted exploration and generally produces better early results because the policy has already seen successful behavior.

### Models and Evaluation

Every training run produces a saved model with full metadata: name, task, algorithm, accuracy, episodes trained, creation date, and size. Model files are stored as artifacts; their metadata is queryable through the Models view and the API.

Any saved model can be evaluated. Evaluation reloads the model, runs it deterministically (choosing consistent actions rather than exploring) across test episodes, and reports:

* Total and average reward
* Success rate (and training vs. target success rate)
* Success gap
* Closeness score (how near the policy came to perfect execution)
* Points awarded

Deterministic evaluation is how you compare models honestly and decide whether a policy is improving.

### The Dataset View

The Dataset view aggregates coverage across all your activity, per task: available task definitions, training sessions, saved models, manual models, episode counts, best accuracy, average accuracy, best reward, and latest activity.

This is one of the most strategically important screens on the platform. It turns collection from guesswork into strategy: if one task has hundreds of episodes at high accuracy while another has none, the view makes the gap obvious and tells the network where to collect next.

### Rewards and Leaderboard

Points are awarded for genuinely useful activity:

* Completing a training run (base reward plus a success-rate bonus)
* High-performance training (extra bonus when final success rate ≥ 80%)
* Successful demonstrations (more than attempted demonstrations)
* Behavior cloning contributions (warm-starting a model from demos)
* Model evaluation (based on closeness to perfect success)

Every point event is written to your rewards ledger. The leaderboard ranks contributors by total points and shows your standing relative to others and to the current leader. See Section 6 for how points convert to revenue shares.


---

# Agent Instructions
This documentation is published with GitBook. GitBook is the documentation platform designed so that both humans and AI agents can read, navigate, and reason over technical content effectively. Learn more at gitbook.com.

## Querying This Documentation
If you need additional information that is not directly available in this page, you can query the documentation dynamically by asking a question.

Perform an HTTP GET request on the current page URL with the `ask` query parameter, and the optional `goal` query parameter:

```
GET https://docs.roboset.org/training-platform/data-training.md?ask=<question>&goal=<endgoal>
```

`ask` is the immediate question: it should be specific, self-contained, and written in natural language.
`goal` is optional and describes the broader end goal you are ultimately trying to accomplish on behalf of the user. GitBook uses it to tailor the answer towards what is most useful for that goal.

The response will contain a direct answer to the question and relevant excerpts and sources from the documentation.

Use this mechanism when the answer is not explicitly present in the current page, you need clarification or additional context, or you want to retrieve related documentation sections.